AI-Enabled Cyberattacks in 2026: Deepfakes, Synthetic Phishing and Shadow AI

Hero Summary

Artificial intelligence affects cybersecurity in two ways: attackers use it to make fraud more convincing and cheaper to scale, and employees' own use of unsanctioned AI tools — "shadow AI" — creates new data-exposure risk inside organizations. The FBI's 2025 Internet Crime Report added an "AI Related" descriptor and recorded 22,364 complaints with reported losses of $893,346,472 carrying that tag. IBM's 2025 research for India found that shadow AI added INR 17.9 million to the average cost of a breach and that only 37% of organizations in India reported having AI access controls in place. The practical response is mostly procedural: verify high-risk requests independently, protect accounts with phishing-resistant authentication, set clear rules for AI use, and train people for a world in which a familiar voice or polished message proves nothing.

Executive Summary

AI has lowered the cost of producing fluent, personalised, error-free messages, cloned voices, and fabricated images and video. That changes one long-standing defensive habit: "this looks and sounds legitimate" is no longer evidence. At the same time, many employees paste company data into public AI tools or use unvetted AI browser extensions and connectors, creating exposure that security teams cannot see. This guide summarises what the evidence shows, the main attacker uses of AI, the shadow-AI problem, and a set of controls — verification, authentication, policy, monitoring, and training — that work regardless of how convincing the fake is.

What the data shows {#data}

• FBI IC3, 2025: complaints carrying the "AI Related" descriptor numbered 22,364, with reported losses of $893,346,472. IC3's descriptor shows where AI was involved in a complaint; it does not capture every AI-assisted attack, so it is a floor, not a total.
• IBM, India 2025: shadow AI was a top-three cost driver and added INR 17.9 million to the average cost of a breach, and only 37% of organizations in India reported having AI access controls in place. The average total cost of a breach in India was INR 220 million.
• Underlying fraud types remain large: phishing and spoofing was the most-reported crime type to IC3 in 2025 (191,561 complaints), and business email compromise losses were $3,046,598,558. AI does not create these crimes; it makes them more convincing.

How attackers use AI {#attackers}

• Drafting persuasive messages in any language, free of the spelling and grammar errors that once gave phishing away.
• Personalising at scale — using public information about a person or company to tailor a lure.
• Cloning voices from short audio samples taken from public videos, calls, or social media.
• Creating deepfake images and video for impersonation, fake identities, and fraudulent job applicants or customers.
• Automating reconnaissance and routine tasks, reducing the skill and time needed to run campaigns.
• Generating fake documents, invoices, and websites quickly.

Deepfakes and cloned voices {#deepfakes}

A cloned voice can make a call from a "CEO" or "bank officer" sound authentic, and a deepfake video call can show a convincing face. These are used for urgent payment requests, credential resets, and help-desk fraud. The lesson for defenders is that recognising a voice or face is not verification. Identity must be confirmed by process: a call-back to a known number, a code word agreed in advance, a second approver, or a verified channel. See Smishing and Vishing and Business Email Compromise.

Synthetic phishing and impersonation {#phishing}

Phishing messages written with AI can be grammatically perfect, context-aware, and tailored to the recipient's role. Defences that rely on spotting poor writing weaken; those that rely on verification, authentication, and reporting do not. Email authentication, phishing-resistant MFA, and quick reporting of anything unexpected stay effective. See Phishing Attacks: Types, Warning Signs & Prevention and Identity Attacks.

Shadow AI: the risk inside the organization {#shadow}

Shadow AI is the use of AI tools that the organization has not approved or does not control. Typical risks:

• Data leakage: employees paste customer data, source code, contracts, or strategy into public AI services.
• Unvetted tools: browser extensions, meeting-transcription assistants, coding assistants, and connectors that request broad access to email, files, or code repositories.
• Personal accounts used for work, which removes organizational visibility and controls.
• Compliance exposure: personal data processed by an unapproved tool or vendor may breach privacy obligations. See DPDP Act Compliance Checklist and Third-Party and Vendor Risk Management Under DPDP.
• Over-trust: people rely on AI output without verification, inserting errors or insecure code into work.

IBM's India findings suggest the cost is real and the controls are lagging: shadow AI added INR 17.9 million to average breach costs, and only 37% of organizations reported AI access controls.

Controls that work {#controls}

Risk	Controls
Impersonation by voice or video	Out-of-band verification, code words, dual approval for payments, call-back procedures
Convincing phishing	Phishing-resistant MFA, email authentication (SPF, DKIM, DMARC), filtering, easy reporting
Account takeover	Number matching or phishing-resistant MFA, short sessions, monitoring
Help-desk fraud	Strict identity verification for resets and MFA changes
Shadow AI data leakage	Approved-tool list, data classification and handling rules, technical controls to block or monitor unsanctioned services, access controls
Over-trust in AI output	Review and testing requirements for AI-generated content and code
Third-party AI risk	Vendor assessment, contract terms on data use and retention, review of connected apps and permissions

An AI-use policy checklist {#policy}

• Which AI tools are approved, for what purposes, and how to request a new one.
• What data may never be entered into external AI tools — personal data, credentials, confidential contracts, source code, unreleased financials.
• Whether employees may use personal accounts for work (generally no).
• Who owns AI risk and approves tools — typically security, legal, and the business together.
• Review and approval requirements for AI-generated outputs used in customer-facing or critical work.
• Logging and monitoring expectations for approved tools.
• Vendor and contract requirements — data location, retention, use for model training, breach notification.
• Training and reporting — how to ask questions and report misuse without fear of blame.

Training for an AI-era threat landscape {#training}

Update training to match reality: teach that polish is not proof, that voices and faces can be faked, and that the safe habit is to verify through a separate trusted channel. Include simulation of AI-style lures across email, text, and voice, and role-based sessions for finance, executives, and help-desk staff. See Role-Based Security Awareness Training and Adaptive Security Awareness Training.

How Threat ResQ fits {#threat-resq-fit}

TRISA is Threat ResQ's AI security awareness and human risk intelligence platform, with an AI mentor that adapts each employee's training path and continuous readiness scoring with department-level and board-level reporting. TRAP, its phishing simulation capability, runs phishing, smishing, and vishing simulation. Threat ResQ's Security Awareness Services provide a facilitated, human-led layer for groups such as executives and finance teams.

FAQ {#faq}

How are attackers using AI? To write convincing, personalised messages, clone voices from short audio samples, create deepfake images and video, automate reconnaissance, and generate fake documents and websites quickly.

How big is AI-enabled fraud? The FBI's IC3 recorded 22,364 complaints and $893,346,472 in reported losses carrying its "AI Related" descriptor in 2025. This is a floor, because not every AI-assisted attack is identified as such.

What is shadow AI? The use of AI tools that the organization has not approved or does not control. IBM's 2025 India research found it added INR 17.9 million to the average cost of a breach.

Can I trust a voice or video call from someone I know? Not for high-risk requests. Voice cloning and deepfakes can imitate people, so verify requests for money, credentials, or access through an independent channel such as a call-back to a known number.

What is the most effective defence against AI-enabled attacks? Process and authentication rather than detection of fakes: independent verification of high-risk requests, phishing-resistant MFA, email authentication, clear AI-use policies, and training that teaches people not to treat polish or familiarity as proof.